gstack
The most engineered skill suite GearScope has tested, and the heaviest.
gstack is a full software factory for Claude Code: 55 skills covering the think-plan-build-review-test-ship loop, 91 helper binaries, 10 host integrations, and a 979-file test suite, all installed and exercised on this machine at v1.91.6.0. The engineering discipline is the best in the tier (hash-chained egress receipts, working-tree fingerprints, a 134-token always-on floor for the whole suite), but you pay for it: 1.1 GB on disk after setup, zero git tags to pin against a repo that shipped four versions in one day, and a stale "23 tools" description in front of 55 shipped skills. If you live in Claude Code and want an enforced process rather than loose prompts, keep it.
$git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup
verified: 82 MB clone, ./setup exit 0, 57 SKILL.md paths registered; needs bun 1.0+
$./setup --host codex\
cursor\|kiro\|factory\|opencode | full installs; hermes, openclaw, and slate get an instruction-only digest tier (verified informational)
install if
- Technical founders and solo builders on Claude Code. The suite's core value is the enforced think-plan-build-review-test-ship loop, and it is the audience the README is written for.
- Small teams standardizing one workflow. Team mode bootstraps a repo so every session auto-updates from one source, with no vendored files.
- Anyone who cares what their agent sends off-machine. The egress receipt ledger with per-consent revoke commands is unique in this tier and works as advertised.
skip if
- Primarily non-Claude users. Hermes, OpenClaw, and Slate get a 2KB instruction digest, not the suite; the full experience is Claude-Code-shaped.
- Minimalists and small-disk machines. 1.1 GB after setup, a Chromium download by default, and 55 skills is a lot of surface for someone who wanted a review prompt.
- Anyone who needs pinned releases. Zero tags across 453 versions means you are always on HEAD of main, with hourly auto-update in team mode.
- Windows users needing the full CSO runtime. The security container's producer path has no Windows build, and several cookie-import and filesystem issues in the tracker are Windows-first.
What It Does
gstack is Garry Tan's (President of Y Combinator) personal Claude Code setup, open sourced: a suite of 55 skills that turns one agent session into a role-based engineering team. Slash commands cover the full sprint: /office-hours and /plan-ceo-review interrogate a product idea, /autoplan chains CEO, design, DX, and eng reviews, /review finds bugs that pass CI, /qa drives a real browser against a staging URL, /cso runs a security audit, /ship and /land-and-deploy carry the PR to production, and /retro closes the loop. Around the conversational skills sit 91 helper binaries in bin/ (working-tree fingerprints, token bills, an egress receipt auditor, evidence ledgers, an iOS QA daemon), a bundled headless Chromium for browser work, and host installers for ten agents (Claude Code, Codex, Cursor, Factory, Kiro, OpenCode, plus digest tiers for OpenClaw, Hermes, and Slate). The target user is a technical founder or solo builder living in Claude Code who wants an enforced process instead of ad hoc prompting. The repo reached 134,454 stars and 20,019 forks (API-verified 2026-09-29) in under seven months, and skills.sh records 39,940 installs across its 57 indexed entries, with the gstack router skill alone at 23,475.
The Good
The verification culture is unlike anything else in the skill tier. The repo carries 979 test files totaling 227,218 lines of test code, with names like agents-digest.test.ts, review-log receipt tests, and browser security-classifier tests. A sample ran green on this machine (6 pass, 0 fail, 23 expect() calls). The changelog holds 453 versioned entries in a documented style (docs/CHANGELOG_STYLE.md), with three releases on 2026-09-28 alone, each entry carrying engineering detail like per-shard wall times. Even the README's productivity claims ship with a methodology doc and a reproduction script (docs/ON_THE_LOC_CONTROVERSY.md). This is the only skill repo GearScope has tested that measures its own claims.
The security engineering is real and it works when you run it. gstack logs every off-machine send to a hash-chained egress ledger before the send happens. On the installed tree, gstack-egress list reported the ledger path, gstack-egress verify returned chain intact: 0 line(s) verified (exit 0, nothing had been sent), and gstack-egress grants listed each standing consent with its revoke command, telemetry defaulting to off and local-only. The same discipline shows up elsewhere: the no-bun failure path prints a checksum-verified bun install procedure, the prompt-injection defense runs a local ML classifier sidecar rather than a network service, tracker text from GitHub issues is routed through a trust envelope (gstack-issue-guard), and NOTICE.md meticulously attributes every line derived from Apache-2.0 works (impeccable's rule catalog, with the exact commits and files).
Token discipline is measured, not claimed. The whole 55-skill suite costs about 134 always-on tokens per session: gstack-context-bill on the installed tree reported 2 always-on skills totaling 536B of frontmatter (_gstack-command at 244B, connect-chrome at 292B), with everything else loaded per invocation. The tool even documents its own estimate error (mean 7.4% per file, bias under 0.5%, calibrated against count_tokens on 219 files) and offers --exact to send content to Anthropic only after writing an egress receipt. Setup also degrades quietly: with gbrain absent it printed that brain-aware blocks are suppressed in planning skills at zero token overhead, exactly as documented.
The wtree fingerprint does what it claims, in both directions. Two consecutive runs of gstack-wtree in a test repo printed the identical hash 42d4c5245460645340a0b5b189f055b93cca0f7e. Adding an untracked source file changed it to 68ba7e4f796cbce5ed86bad3e9df986fb138d99f. Committing that same content left the fingerprint unchanged. That is the content-vs-commit-SHA binding the review receipts depend on, verified with a positive and a negative case.
Multi-host support is tiered and honest about it. The hermes path was tested directly: ./setup --host hermes exits 0 as an informational tier, printing the 2KB digest path (agents-digest/gstack-AGENTS.md) instead of pretending to install. The digest's version line matched the VERSION file (1.91.6.0), and the repo's own test suite asserts that setup never copies the digest onto a user's AGENTS.md uninvited.
The Bad
The install weight is the heaviest GearScope has recorded for a skills package. The shallow clone is 82 MB, of which 34 MB is test fixtures and 20 MB is .git history riding along with the payload. After ./setup, the tree is 1.1 GB with 693 MB of node_modules, and the default setup also downloads a Playwright Chromium build (skipped here only via the documented GSTACK_SKIP_PLAYWRIGHT=1 flag). For comparison, the onetake video skill weighed 71 MB and that was flagged as heavy. A first-time user who pastes the README install gets a browser, two compiled binaries, and a package tree before a single skill runs.
There is no version anchor to pin against. The repo has zero git tags and zero GitHub releases across 453 changelog versions. The documented install clones HEAD of main, /gstack-upgrade is the only update path, and team mode auto-updates hourly. Between this review's first clone (v1.91.2.0, 2026-09-28) and its re-test the next day, main had already moved four versions to v1.91.6.0. Anyone running gstack across a team is accepting continuous upstream drift with no rollback target except a git reflog.
The public description undercounts the shipped surface by more than half. The GitHub description says "23 opinionated tools" and the README hero says "Twenty-three specialists and eight power tools." The repo actually ships 55 skill directories (62 SKILL.md files), the README's own tables list 32 sprint skills plus 13 power tools plus 15 standalone binaries, and the install prompt enumerates 37 commands. The number appears frozen from an earlier era of the project, and it is the first thing a visitor reads.
Support load is real and lands on host setup. There are 371 open true issues (941 open items including PRs). The most-commented threads are recurring install breakage: Conductor template failures with build exit code 128 (#1174), "setup for cursor is broken" (#1269), and a user confused that setup --host hermes is a help banner, not an installer (#1201). The hermes behavior is documented and verified working as designed, but the issue count shows the multi-host surface outrunning the one-person maintenance core (garrytan has 368 of roughly 409 commits).
Smoke Test Results
Host-based runs on macOS 15 aarch64 (the sbx daemon was unavailable), bun 1.3.14, git shallow clones. Run B re-used a tree first installed at v1.91.2.0 and hard-updated to v1.91.6.0 before the run, so ./setup exercised the incremental path (12s vs 32s first run). Logs are the scripted runs, captured verbatim.
Run A. Fresh sandbox, no deps preinstalled
$ git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git
β PASS: clone succeeds per the README install command
$ du -sh gstack
β PASS: 82M measured (34M test fixtures, 20M .git)
$ ./setup --help
β PASS: prints usage including --host without bun installed
$ ./setup (PATH stripped of bun)
β PASS: exit 1, names bun, prints checksum-verified install procedure
$ test -e ~/.claude/skills or ~/.gstack in isolated HOME
β PASS: failed setup left no partial install
Pass rate: 5 of 5. The no-dep failure mode is handled and self-explaining.
Run B. Sandbox with deps preinstalled (bun on host)
$ command -v bun
β PASS: bun 1.3.14 available
$ git clone ... ~/.claude/skills/gstack (isolated HOME)
β PASS: cloned to the README path
$ ./setup (GSTACK_SKIP_PLAYWRIGHT=1 GSTACK_SKIP_COREUTILS=1)
β PASS: exit 0 in 12s, prints first-move guidance (/office-hours, /qa)
$ test -x browse/dist/browse
β PASS: browse binary built and executable
$ test -x make-pdf/dist/pdf
β PASS: make-pdf binary built and executable
$ browse/dist/browse --help
β PASS: responds with the browse usage banner
$ find ~/.claude/skills -name SKILL.md | wc -l
β PASS: 57 SKILL.md paths registered
$ test -e ~/.claude/skills/office-hours/SKILL.md
β PASS: office-hours registered
$ test -e ~/.claude/skills/review/SKILL.md
β PASS: review registered
$ test -e ~/.claude/skills/ship/SKILL.md
β PASS: ship registered
$ test -e ~/.claude/skills/qa/SKILL.md
β PASS: qa registered
$ test -e ~/.claude/skills/cso/SKILL.md
β PASS: cso registered
$ test -e ~/.claude/skills/design-html/SKILL.md
β PASS: design-html registered
$ test -L ~/.claude/skills/review/SKILL.md
β PASS: symlink back into the gstack source (documented unix shape)
$ grep gstack-timeline-stop ~/.claude/settings.json
β PASS: default-on Stop hook registered, with the removal command printed
$ du -sh the installed tree
β PASS: repo 1.1G after build, node_modules 693M
Pass rate: 16 of 16. The documented install path works end to end; the weight finding above is the cost.
Run C. Functional verification (does it do what it claims?)
$ gstack-wtree (run twice in a test repo)
β PASS: identical fingerprints 42d4c5245460645340a0b5b189f055b93cca0f7e
$ gstack-wtree after adding an untracked source file
β PASS: fingerprint changes to 68ba7e4f796cbce5ed86bad3e9df986fb138d99f
$ gstack-wtree after committing that same content
β PASS: fingerprint unchanged (content binding, not SHA binding)
$ gstack-context-bill ~/.claude/skills
β PASS: numeric bill; 2 always-on skills, 536B (~134 tok) for the whole suite
$ head -5 agents-digest/gstack-AGENTS.md vs VERSION
β PASS: digest version 1.91.6.0 matches the VERSION file
$ ./setup --host hermes
β PASS: exit 0 as an informational tier, prints digest path, installs nothing
$ bun test test/agents-digest.test.ts
β PASS: 6 pass, 0 fail, 23 expect() calls, including "setup never copies the digest onto a user AGENTS.md"
Functional pass rate: 7 of 7.
What the runs tell you
The install path and the tool-level claims (fingerprints, token bills, digest sync, egress ledger) all hold under assertion. What this harness cannot reach is the in-session experience: /review finding a real bug or /qa driving a browser requires a live Claude Code session with API access, so the suite's headline value is evidenced by its own test suite and registry demand rather than by direct execution here.
Setup Walkthrough
- Install bun (the one hard dependency; the setup script prints a checksum-verified install command if it is missing).
- Clone per the README:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup. First run took 32s here; it builds the browse and make-pdf binaries and registers 57 SKILL.md paths. - Read the setup tail output. It tells you what was skipped (Chromium if you set GSTACK_SKIP_PLAYWRIGHT=1, plan-tune hooks in non-interactive runs) and prints the exact commands to add them later.
- Note the two things setup changes outside its directory: a Stop hook (gstack-timeline-stop, fail-open, 2s budget) in ~/.claude/settings.json, and per-repo hooks only if you later opt in. The removal commands are printed in the setup output and gstack-uninstall exists.
- Other agents:
./setup --host codex|cursor|kiro|factory|opencodefor full installs, or copy the 2KB digest from agents-digest/gstack-AGENTS.md for any rules-reading agent.
Post-install gotcha: there is no release tag to pin, so note the VERSION file (1.91.6.0 here) if you need to reference what you are running.
Alternatives
- obra/superpowers - the 292K-star methodology pack: craft-level skills (TDD, planning, debugging) you adopt piecemeal, rather than one opinionated team pipeline with binaries.
- Donchitos/Claude-Code-Game-Studios - a 49-agent, 74-skill studio template for game development; the same whole-team philosophy in a domain-specific, template-driven form.
- forrestchang/andrej-karpathy-skills - rules-only guardrails against the four failure modes gstack's workflow enforces; the lightweight option gstack itself positions against.
- anthropics/skills - the official Anthropic skill pack: per-task skills (documents, artifacts) with no process opinion, for when you want primitives instead of a process.
Reviews stay honest because nobody pays us to publish them. If this one saved you time, throw a coin.
Tip the reviewer- reviewed by
- GearScope
- tested
- 2026-09-29 · macOS (Apple Silicon)
- last verified
- 2026-09-29
- depth
- HANDS-ON
- sponsorship
- none, ever
Don't install your next skill blind.
Every week: the shortlist of skills worth installing β and the ones to skip β from 100+ hands-on tests. No spam, no affiliate links.